Cloud-based pension information management system

Abstract

The invention relates to the technical field of pension information management, and discloses a cloud-based pension information management system, which comprises a health data acquisition module, a privacy protection processing module, a risk prediction module, a resource scheduling module and a comprehensive service module. The health data acquisition module synchronously acquires data in a multi-mode manner and encrypts and uploads the data by means of wearable equipment and an Internet of Things sensor; the privacy protection processing module guarantees data security by using federated learning and differential privacy encryption; the risk prediction module accurately predicts health risks based on a graph neural network; the resource scheduling module optimizes resource configuration through deep reinforcement learning; and the comprehensive service module generates a personalized service chain according to the knowledge graph. The system solves many problems of traditional pension information management, realizes multi-dimensional data real-time acquisition and privacy protection, accurately predicts health risks, efficiently schedules resources, provides personalized services, improves the pension service quality and the life quality of old people, and promotes the intelligent development of the pension industry.

Description

Technical Field

[0001] The present invention relates to the technical field of elderly care information management, and in particular to a cloud-based elderly care information management system. Background Art

[0002] As the global population ages, the issue of elderly care has attracted much attention. The traditional elderly care model has prominent problems in information management and service supply, and the application of information technology in the elderly care field also has many shortcomings.

[0003] Traditional methods rely on regular physical examinations and manual records. Data collection is one-sided, delayed, and prone to errors. In addition, it is impossible to obtain multi-dimensional data, making it difficult to fully assess the health status of the elderly. At the same time, there is a lack of means to protect data privacy, and there is a lack of encryption and desensitization mechanisms when storing and sharing, which poses a risk of leakage. Relying on experience to judge health risks, it is impossible to accurately assess them, and ignoring environmental and social factors makes it difficult to prevent diseases and accidents in advance. Resource scheduling relies on manual experience and lacks flexibility, resulting in uneven distribution and untimely resource allocation in emergencies. Traditional elderly care services are standardized and do not take into account individual differences among the elderly. They cannot meet diverse needs in terms of rehabilitation care, social activities, etc., which reduces the quality of life and satisfaction of the elderly.

[0004] At present, cloud computing is mostly used for data storage in the elderly care field. Its powerful computing and scheduling capabilities have not been fully tapped, and it is impossible to analyze large-scale data in real time, making it difficult to support the optimization of elderly care services. There are many IoT devices in the elderly care environment, but the communication protocols and data formats of devices from different manufacturers are not unified, which makes it difficult for devices to collaborate, forming "information islands" and affecting the comprehensive perception of the life and health status of the elderly.

[0005] In the management of elderly care information, the application of artificial intelligence algorithms is still shallow. The health risk prediction model is single and the prediction accuracy is poor; the service recommendation algorithm cannot accurately match the needs of the elderly, and the application value is limited. The existing elderly care information management system is mostly a patchwork of modules, lacking an overall design, and insufficient data interaction and business process integration among modules, which affects the efficiency and quality of elderly care services. Summary of the invention

[0006] The purpose of the present invention is to provide a cloud-based elderly care information management system to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a cloud-based elderly care information management system, including a health data collection module, a privacy protection processing module, a risk prediction module, a resource scheduling module and a comprehensive service module; The health data collection module is used to collect multi-dimensional health data of elderly users, specifically by deploying a wearable device cluster and an IoT sensor to collect data in real time, obtain the user's comprehensive health raw data, and encrypt the user's comprehensive health raw data and upload it to the privacy protection processing module; The real-time data collection specifically refers to the synchronous collection of multimodal data, including collecting physiological index data once every 30 seconds, collecting activity trajectory data once every 5 minutes, and collecting environmental safety data once every hour, and obtaining the user's comprehensive health raw data through the synchronous collection of multimodal data, and storing the user's comprehensive health raw data in the edge computing node; The privacy protection processing module is used to anonymize and enhance the security of the original data, specifically to perform distributed feature extraction and differential privacy encryption on the user's comprehensive health original data through a federated learning framework to obtain desensitized health feature data, and send the desensitized health feature data to the risk prediction module and the resource scheduling module. The specific steps include: data sharding processing, federated feature aggregation, dynamic noise injection and cross-domain privacy verification; The risk prediction module is used to analyze the desensitized health feature data and predict the user's health risk level, specifically to generate a health risk prediction result using a spatiotemporal joint modeling method based on a graph neural network, and send the health risk prediction result to the resource scheduling module; The resource scheduling module is used to generate personalized service solutions and dynamically optimize resource allocation strategies through a deep reinforcement learning framework based on the health risk prediction results and the desensitized health feature data, and output resource scheduling instructions to the comprehensive service module; The comprehensive service module is used to integrate the resource scheduling instructions and real-time user needs, generate a personalized service chain using a knowledge graph-driven service recommendation method, and execute service instructions through a voice interaction interface to complete closed-loop management of elderly care services.

[0008] Preferably, the step of obtaining the health risk prediction result by the risk prediction module includes: constructing a spatiotemporal graph structure operator, constructing a node relationship reinforcement operator, constructing a multi-scale graph convolution sub-model, constructing a temporal attention sub-model and constructing a risk prediction integrated model; The said construction of spatiotemporal graph structure operator is used to map the user's physiological indicators, social relationships and environmental factors into dynamic graph nodes and edge weights, specifically using a spatiotemporal embedding algorithm to generate a multi-dimensional graph topology structure; The node relationship strengthening operator is constructed, specifically by calculating the association strength between nodes through a graph attention mechanism, and optimizing the edge weight distribution using a relationship graph transformer; The multi-scale graph convolution sub-model is constructed, specifically, a layered graph convolution network is designed to extract local physical sign association features and global social influence features respectively, and a joint graph feature matrix is ​​generated through a feature fusion layer; The construction of the temporal attention sub-model specifically adopts a bidirectional temporal attention mechanism to mine periodic patterns of historical health data and output a temporal feature vector; The construction of the risk prediction integrated model specifically involves performing tensor splicing on the joint graph feature matrix and the time series feature vector, and classifying the risk levels through a multi-layer perceptron to obtain a health risk prediction result.

[0009] Preferably, in the privacy protection processing module, the federated feature aggregation specifically adopts an asynchronous gradient update strategy, generates feature embedding vectors through edge node local model training, and performs weighted aggregation in the cloud using a homomorphic encryption algorithm; The dynamic noise injection is specifically to dynamically adjust the Laplace noise intensity according to the data sensitivity. The noise injection formula is:

[0010] in, express The noise intensity at the moment, is the privacy budget coefficient, is the failure probability threshold, is the information entropy value of the current data slice.

[0011] Preferably, the resource scheduling module uses a deep reinforcement learning framework to make dynamic decisions based on the desensitized health feature data and the health risk prediction results, specifically including: defining state space, building a double-delayed deep deterministic policy gradient network, designing reward functions and online policy optimization; The defined state space specifically encodes the user health status, service resource inventory, environmental safety index and device availability into a multi-dimensional state vector; The dual-delayed deep deterministic policy gradient network is constructed, specifically using a master-slave dual critic architecture, generating resource configuration actions through a policy network, and achieving training stability through delayed updates of a target network; The design reward function is specifically to construct a composite reward function based on service response efficiency, resource utilization and risk mitigation degree:

[0012] in, Score the response efficiency. is the resource utilization rate, is the unresolved risk weight, , , It is a dynamic adjustment coefficient.

[0013] Preferably, in the comprehensive service module, the service recommendation method driven by the knowledge graph comprises the following specific steps: constructing an ontology library in the field of elderly care, generating a dynamic knowledge graph, performing multimodal semantic matching, and outputting a personalized service chain; The construction of the elderly care ontology library specifically defines six entity types, including user portraits, medical resources, nursing services, social activities, safety equipment, and emergency strategies, and establishes 53 types of associations between entities; Generating a dynamic knowledge graph specifically involves incrementally updating the real-time collected environmental data, user behavior data, and service record data through a graph database to form a dynamic knowledge graph with a timestamp; The multimodal semantic matching is performed by adopting a cross-modal transformer model to jointly encode user voice requests, text work orders and sensor data to generate a unified semantic vector; The output personalized service chain specifically performs multi-hop reasoning in the knowledge graph based on semantic vectors to generate a comprehensive service sequence including medical service appointments, activity recommendations, and safety warning linkage.

[0014] Preferably, in the health data collection module, the user's comprehensive health raw data specifically includes: heart rate variability data, blood oxygen saturation data, gait feature data, indoor positioning trajectory data, air quality index data and emergency call event record data; The edge computing node deploys a lightweight data preprocessing unit to specifically perform data format standardization, outlier filtering and feature dimension reduction operations.

[0015] Preferably, the health risk prediction results output by the risk prediction module specifically include: disease onset probability value, fall risk index, loneliness tendency score and nutritional imbalance warning level; The disease onset probability value is calculated by analyzing 72 hours of continuous physiological data through a time series convolutional network, and the fall risk index is generated by combining gait characteristics and environmental obstacle distribution through graph neural network reasoning.

[0016] Preferably, the resource scheduling module is configured with an elastic resource pool, which includes three types of schedulable resources: a mobile nursing robot cluster, an emergency medical supplies inventory, and an online expert consultation channel; The deep reinforcement learning framework performs a strategy update every 15 minutes, dynamically adjusts the allocation ratio of the three types of resources according to real-time demand changes, and verifies the feasibility of the scheduling plan through digital twin system simulation.

[0017] Preferably, the integrated service module integrates a voice interaction interface, specifically adopts an end-to-end speech synthesis and recognition model, supports dialect adaptation and fuzzy instruction parsing; The voice interaction interface has a built-in anti-noise enhancement unit, which eliminates environmental noise interference through a generative adversarial network, ensuring that the command recognition accuracy rate is maintained above 90% under 60 decibel background noise.

[0018] Compared with the prior art, the present invention has the following beneficial effects: With the help of wearable device clusters and IoT sensors, the system realizes synchronous high-frequency collection of multimodal data. Physiological index data is collected every 30 seconds, which can capture the instantaneous changes of key physiological parameters of the elderly such as heart rate and blood oxygen saturation in real time; activity trajectory data is collected every 5 minutes to detect abnormal behavior of the elderly in time; environmental safety data is collected every hour to monitor the safety of the living environment at all times. This breaks through the limitations of traditional regular physical examinations and manual records, and provides rich and accurate data support for a comprehensive and dynamic grasp of the health status of the elderly.

[0019] The federated learning framework and differential privacy encryption technology are used to process the user's comprehensive health raw data in multiple links. Data sharding processes scattered data to reduce the risk of leakage; federated feature aggregation realizes feature fusion under the premise of ensuring data security; dynamic noise injection adjusts the noise intensity according to the data sensitivity to further confuse the data; cross-domain privacy verification ensures the security of data when interacting in different regions. It effectively protects the sensitive information of the elderly, allowing them to use the system with confidence and avoid resisting digital elderly care services due to privacy concerns. The spatiotemporal joint modeling method based on graph neural networks comprehensively considers physiological indicators, social relationships and environmental factors. By constructing a variety of operators and sub-models, the deep correlation of data is mined. The probability value of disease onset is obtained by analyzing 72 hours of continuous physiological data through a time series convolutional network, and the fall risk index is generated by combining gait and environmental obstacle distribution inference. Compared with the traditional simple judgment based on physiological indicators or experience, the prediction results are more accurate and can detect potential health risks in advance.

[0020] Accurate risk prediction makes early intervention possible. For elderly people who are predicted to be at risk of falling, protective facilities can be set up in their living environment in advance and special personnel can be arranged to pay attention to them; for elderly people with a high probability of disease onset, care plans can be adjusted in a timely manner and medical treatment can be reminded. This reduces the incidence of health risk events, reduces the physical pain and medical expenses of the elderly, and ensures their quality of life and life safety.

[0021] The resource scheduling module uses a deep reinforcement learning framework to make dynamic decisions based on multiple factors. The user's health status, resource inventory, environmental safety index and equipment availability are encoded as state vectors, a stable network architecture is constructed, and a composite reward function is designed. The resource allocation ratio of mobile nursing robots, emergency medical supplies and online expert consultation channels is adjusted every 15 minutes according to real-time needs to improve resource utilization efficiency and avoid the coexistence of resource waste and shortage. The knowledge graph of the comprehensive service module drives the service recommendation method, builds an ontology library and a dynamic knowledge graph, and performs multimodal semantic matching. It can generate a personalized service chain based on the unique needs of each elderly person. It arranges precision medical services for elderly people with chronic diseases and recommends social activities for lonely elderly people. It improves the pertinence and effectiveness of services and enhances the satisfaction and sense of gain of elderly people with elderly care services. The voice interaction interface of the comprehensive service module adopts an advanced model, supports dialect adaptation and fuzzy command parsing, and has a built-in anti-noise enhancement unit. Elderly people can easily issue commands through voice without complex operations, and can accurately recognize them even in noisy environments. It improves the convenience and autonomy of elderly people's interaction with the system, and facilitates the use of elderly people who are not familiar with digital devices.

[0022] The modules of the system work closely together to form a closed-loop management. Health data collection provides a basis for risk prediction, risk prediction results guide resource scheduling, resource scheduling instructions drive comprehensive service modules, and comprehensive service feedback can optimize the entire process. This integrated model improves the overall efficiency of elderly care services, reduces information transmission delays and errors, promotes the intelligent upgrading of the elderly care industry, and has good social and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a working principle diagram of the cloud-based elderly care information management system of the present invention; Figure 2 Working principle diagrams generated for risk prediction and key indicators; Figure 3 A working diagram for resource scheduling and service recommendations. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] See also Figure 1-3 The present invention provides a technical solution: a cloud-based pension information management system, the system comprising: Health data collection module: By deploying wearable device clusters and IoT sensors, the health data of elderly care users is collected in real time in multiple dimensions. The synchronous collection of multimodal data is specifically manifested in collecting physiological index data such as heart rate variability data and blood oxygen saturation data every 30 seconds, collecting activity trajectory data such as indoor positioning trajectory data every 5 minutes, and collecting environmental safety data such as air quality index data every hour. The collected user comprehensive health raw data includes heart rate variability data, blood oxygen saturation data, gait feature data, indoor positioning trajectory data, air quality index data, and emergency call event record data. After collection, the data is first stored in the edge computing node, which is deployed with a lightweight data preprocessing unit, which will perform data format standardization, outlier filtering, and feature dimension reduction operations, and then encrypt the processed raw data and upload it to the privacy protection processing module.

[0026] Privacy protection processing module: Receives the user's comprehensive health raw data from the health data collection module, and performs distributed feature extraction and differential privacy encryption on it through the federated learning framework. The specific steps include data sharding, federated feature aggregation, dynamic noise injection, and cross-domain privacy verification. After processing, desensitized health feature data is obtained and sent to the risk prediction module and resource scheduling module.

[0027] Risk prediction module: Obtain the desensitized health feature data from the privacy protection processing module, and use the spatiotemporal joint modeling method based on graph neural network to generate health risk prediction results. The specific steps involve constructing a spatiotemporal graph structure operator, a node relationship reinforcement operator, a multi-scale graph convolution sub-model, a temporal attention sub-model, and a risk prediction integrated model. The generated health risk prediction results include disease onset probability values, fall risk indexes, loneliness tendency scores, and nutritional imbalance warning levels, and the results are sent to the resource scheduling module.

[0028] Resource scheduling module: Based on the desensitized health feature data and health risk prediction results, a deep reinforcement learning framework is used to generate personalized service solutions and dynamically optimize resource allocation strategies. Specifically, it includes defining the state space, building a double-delayed deep deterministic policy gradient network, designing reward functions, and online policy optimization. Then, the resource scheduling instructions are output to the comprehensive service module.

[0029] Comprehensive service module: Integrates resource scheduling instructions and real-time user needs, uses knowledge graph-driven service recommendation methods to generate personalized service chains, and executes service instructions through voice interaction interfaces to complete closed-loop management of elderly care services.

[0030] The present invention will be further described below in conjunction with Examples 1 to 5: Embodiment 1: This embodiment mainly describes the specific implementation method of the risk prediction module to obtain health risk prediction results. The risk prediction module plays a key role in the entire elderly care information management system. It predicts the user's health risk level by analyzing the desensitized health feature data. The specific implementation steps include: Constructing spatiotemporal graph structure operators: Using spatiotemporal embedding algorithms, users’ physiological indicators (such as heart rate variability, blood oxygen saturation, etc.), social relationships (such as the frequency of interaction with family, friends, and medical staff), and environmental factors (such as indoor temperature, air quality index, etc.) are mapped into dynamic graph nodes and edge weights to generate a multi-dimensional graph topology. In this way, various types of user information can be structured in the form of a graph for easy subsequent analysis.

[0031] Construct node relationship reinforcement operator: Use graph attention mechanism to calculate the strength of association between nodes, such as calculating the degree of association between physiological indicator nodes and environmental factor nodes. At the same time, use relationship graph transformer to optimize edge weight distribution so that the relationship between nodes can more accurately reflect the actual situation.

[0032] Construct a multi-scale graph convolution sub-model: Design a hierarchical graph convolution network, where the bottom layer extracts local vital sign association features, such as extracting association features between heart rate variability data and blood oxygen saturation data; the high-level network extracts global social influence features, such as analyzing the impact of social relationships on user health. Then, the local vital sign association features and global social influence features are fused through the feature fusion layer to generate a joint graph feature matrix.

[0033] Constructing a temporal attention sub-model: Using a bidirectional temporal attention mechanism to mine periodic patterns in historical health data. For example, by analyzing the user's heart rate data over a period of time, we can mine whether there is a specific periodic change pattern. After mining, we output a temporal feature vector that contains key information in the historical health data.

[0034] Constructing an integrated risk prediction model: tensor splicing the joint graph feature matrix and the time series feature vector to fuse the features of the spatial and temporal dimensions. Then, risk level classification is performed through a multi-layer perceptron to finally obtain health risk prediction results, including disease onset probability value, fall risk index, loneliness tendency score, and nutritional imbalance warning level. Among them, the disease onset probability value is calculated by analyzing 72 hours of continuous physiological data through a time series convolutional network, and the fall risk index is generated by combining gait characteristics and environmental obstacle distribution through graph neural network reasoning.

[0035] Embodiment 2: This embodiment is used to describe the federated feature aggregation and dynamic noise injection operations in the privacy protection processing module. These two operations are crucial to protecting user data privacy and ensuring that user information is not leaked during data processing and transmission.

[0036] Using an asynchronous gradient update strategy, edge nodes use locally stored user comprehensive health raw data to perform local model training and generate feature embedding vectors. For example, on an edge node, the collected health data of a user over a period of time is trained to extract feature vectors related to the user's health status. Then, weighted aggregation is performed in the cloud using the homomorphic encryption algorithm. The homomorphic encryption algorithm ensures the security of computing on encrypted data, allowing the cloud to perform weighted aggregation on the feature embedding vectors of each edge node without decrypting the data to obtain a comprehensive feature vector.

[0037] The Laplace noise intensity is dynamically adjusted according to the data sensitivity, and the noise injection formula used is: .in, express The noise intensity at the moment, The privacy budget coefficient is set according to the system's privacy protection requirements and data sensitivity. For example, for data involving sensitive medical information of users, the coefficient can be appropriately reduced to enhance privacy protection. is the failure probability threshold, which is generally set to a smaller value, such as 0.001, indicating the maximum failure probability allowed during the noise injection process; The information entropy value of the current data slice reflects the uncertainty of the data. The more complex the data is and the higher the uncertainty is, the greater the information entropy value is, and the greater the noise intensity is. Through dynamic noise injection, while ensuring data availability, data privacy protection is further enhanced.

[0038] Embodiment 3: This embodiment elaborates on the dynamic decision-making process of the resource scheduling module based on the deep reinforcement learning framework, which helps the system to reasonably allocate resources according to the actual situation of users and improve service quality and resource utilization efficiency. The specific method includes: Define the state space: Encode the user's health status (such as the probability of disease onset, fall risk index and other health risk prediction results), service resource inventory (the number of mobile nursing robot clusters, the types and quantities of emergency medical supplies inventory, etc.), environmental safety index (the degree of safety reflected by environmental data such as air quality index and indoor temperature), and device availability (the working status of wearable devices, IoT sensors and other devices) into a multidimensional state vector. For example, encode the user's current disease onset probability value of 0.2, the availability of 3 mobile nursing robots, the inventory of a certain key drug in emergency medical supplies is 50 units, the air quality index is good, and all sensors are working properly into a multidimensional vector as the input state of the deep reinforcement learning model.

[0039] Constructing a double-delayed deep deterministic policy gradient network: Using a master-slave dual critic architecture, the policy network generates resource allocation actions based on the input state vector, such as deciding how many mobile nursing robots to send to serve a user, how much emergency medical supplies to allocate, etc. The target network is updated with a delay, and in this way, training stability is achieved, avoiding excessive fluctuations in the model during training, so that the model can learn the optimal resource allocation strategy more accurately.

[0040] Design reward function: Construct a compound reward function based on service response efficiency, resource utilization, and risk mitigation: .in, Score response efficiency, for example, the shorter the time from receiving a service request to starting to perform the service, the higher the score; The higher the ratio of actual resource usage to total resource usage, the higher the resource utilization. It is the unmitigated risk weight, which is the weight corresponding to the risk that has not been effectively mitigated in the risk prediction results. The higher the weight, the greater the risk. , , To dynamically adjust the coefficient, adjust it according to different scenarios and needs. For example, in emergency medical scenarios, the coefficient can be appropriately increased. value, and pays more attention to the degree of risk mitigation.

[0041] Online strategy optimization: The deep reinforcement learning framework performs a strategy update every 15 minutes, dynamically adjusting the allocation ratio of three types of dispatchable resources: mobile nursing robot clusters, emergency medical supplies inventory, and online expert consultation channels according to real-time demand changes. By continuously optimizing the strategy, resource scheduling is more in line with actual needs, and the feasibility of the scheduling plan is verified through digital twin system simulation to avoid unreasonable resource allocation in actual applications.

[0042] Embodiment 4: This embodiment describes in detail a service recommendation method driven by a knowledge graph in a comprehensive service module. This method can generate a personalized service chain for users based on their real-time needs and system resource conditions, thereby improving the accuracy and satisfaction of elderly care services. Specific implementation methods include: Constructing an ontology library for the elderly care sector: six entity types are defined, including user portraits (including basic user information, health status, hobbies, etc.), medical resources (hospitals, doctors, medicines, etc.), nursing services (daily care, rehabilitation care, etc.), social activities (senior activity center activities, community gatherings, etc.), safety equipment (smoke alarms, fall alarms, etc.) and emergency strategies (processes for different emergency situations), and 53 associations between entities are established. For example, the health status in the user portrait is associated with the doctor's expertise in the medical resources, so that appropriate medical services can be recommended to users.

[0043] Generate dynamic knowledge graph: Incrementally update the real-time collected environmental data (such as air quality index, temperature change, etc.), user behavior data (activity trajectory, service usage records, etc.) and service record data through the graph database to form a dynamic knowledge graph with timestamp. For example, when a user uses a rehabilitation care service, the service record and the corresponding time information are updated to the knowledge graph, so that the knowledge graph can reflect the user's latest situation in real time.

[0044] Perform multimodal semantic matching: Use a cross-modal transformer model to jointly encode user voice requests, text work orders, and sensor data to generate a unified semantic vector. For example, when a user makes a voice request "I need someone to chat with me", and the sensor detects that the user is currently lonely, the cross-modal transformer model jointly encodes this information and converts it into a semantic vector that can represent the user's needs.

[0045] Output personalized service chain: Based on semantic vectors, multi-hop reasoning is performed in the knowledge graph to generate a comprehensive service sequence including medical service appointments, activity recommendations, and safety warning linkage. For example, if the semantic vector indicates that the user is at risk of falling and has not had a physical examination recently, the system will recommend making an appointment for medical services for relevant examinations, recommend some suitable rehabilitation activities, and link safety equipment to set up fall warnings, forming a personalized service chain.

[0046] Embodiment 5: This embodiment describes in detail some supplementary features of each module of the system, which further improve the functions of the cloud-based elderly care information management system and enhance the practicality and reliability of the system.

[0047] Health data collection module: In addition to common physiological indicators, activity trajectories, and environmental safety data, the user's comprehensive health raw data also includes emergency call event record data to provide a more comprehensive understanding of the user's health status and emergency situations. The lightweight data preprocessing unit deployed on the edge computing node performs data format standardization, outlier filtering, and feature dimension reduction operations. For example, it converts data in various formats collected by different sensors into a format that the system can recognize and process, removes data points that are obviously erroneous or abnormal, and performs feature dimension reduction on the data to reduce the amount of data processing and improve system operation efficiency.

[0048] Risk prediction module: The output health risk prediction results include the probability of disease onset, fall risk index, loneliness tendency score and nutritional imbalance warning level. The probability of disease onset is calculated by analyzing 72 hours of continuous physiological data through a time series convolutional network. This process extracts the time series features in the data through the convolution operation of continuous physiological data, thereby predicting the possibility of disease onset. The fall risk index is generated by combining gait characteristics and environmental obstacle distribution through graph neural network reasoning, which fully considers the impact of the user's own movement state and surrounding environmental factors on the risk of falling.

[0049] Resource scheduling module: Configure an elastic resource pool, including three types of schedulable resources: mobile nursing robot clusters, emergency medical supplies inventory, and online expert consultation channels. The deep reinforcement learning framework performs a policy update every 15 minutes, dynamically adjusts the allocation ratio of the three types of resources according to real-time demand changes, and verifies the feasibility of the scheduling plan through digital twin system simulation. The digital twin system builds a virtual model corresponding to the actual resources and user conditions, simulates the scheduling plan in this virtual environment, discovers possible problems in advance and optimizes them.

[0050] Comprehensive service module: integrated voice interaction interface, using end-to-end speech synthesis and recognition model, supporting dialect adaptation and fuzzy command parsing. For example, when elderly people from different regions make service requests in dialects, the voice interaction interface can accurately identify and convert them into commands that the system can understand. The voice interaction interface has a built-in anti-noise enhancement unit, which eliminates environmental noise interference through the generation of adversarial networks, ensuring that the command recognition accuracy rate is maintained at more than 90% under 60 decibel background noise, ensuring effective communication between users and the system in complex environments.

[0051] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0052] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A cloud-based elderly care information management system, characterized by: It includes health data collection module, privacy protection processing module, risk prediction module, resource scheduling module and comprehensive service module; The health data collection module is used to collect multi-dimensional health data of elderly users, specifically by deploying a wearable device cluster and an IoT sensor to collect data in real time, obtain the user's comprehensive health raw data, and encrypt the user's comprehensive health raw data and upload it to the privacy protection processing module; The real-time data collection specifically refers to the synchronous collection of multimodal data, including collecting physiological index data once every 30 seconds, collecting activity trajectory data once every 5 minutes, and collecting environmental safety data once every hour, and obtaining the user's comprehensive health raw data through the synchronous collection of multimodal data, and storing the user's comprehensive health raw data in the edge computing node; The privacy protection processing module is used to anonymize and enhance the security of the original data, specifically to perform distributed feature extraction and differential privacy encryption on the user's comprehensive health original data through a federated learning framework to obtain desensitized health feature data, and send the desensitized health feature data to the risk prediction module and the resource scheduling module. The specific steps include: data sharding processing, federated feature aggregation, dynamic noise injection and cross-domain privacy verification; The risk prediction module is used to analyze the desensitized health feature data and predict the user's health risk level, specifically to generate a health risk prediction result using a spatiotemporal joint modeling method based on a graph neural network, and send the health risk prediction result to the resource scheduling module; The resource scheduling module is used to generate personalized service solutions and dynamically optimize resource allocation strategies through a deep reinforcement learning framework based on the health risk prediction results and the desensitized health feature data, and output resource scheduling instructions to the comprehensive service module; The comprehensive service module is used to integrate the resource scheduling instructions and real-time user needs, generate a personalized service chain using a knowledge graph-driven service recommendation method, and execute service instructions through a voice interaction interface to complete closed-loop management of elderly care services.

2. The cloud-based elderly care information management system according to claim 1, characterized in that: The steps of obtaining the health risk prediction results by the risk prediction module include: constructing a spatiotemporal graph structure operator, constructing a node relationship reinforcement operator, constructing a multi-scale graph convolution sub-model, constructing a temporal attention sub-model, and constructing a risk prediction integrated model; The said construction of spatiotemporal graph structure operator is used to map the user's physiological indicators, social relationships and environmental factors into dynamic graph nodes and edge weights, specifically using a spatiotemporal embedding algorithm to generate a multi-dimensional graph topology structure; The node relationship strengthening operator is constructed, specifically by calculating the association strength between nodes through a graph attention mechanism, and optimizing the edge weight distribution using a relationship graph transformer; The multi-scale graph convolution sub-model is constructed, specifically, a layered graph convolution network is designed to extract local physical sign association features and global social influence features respectively, and a joint graph feature matrix is ​​generated through a feature fusion layer; The construction of the temporal attention sub-model specifically adopts a bidirectional temporal attention mechanism to mine periodic patterns of historical health data and output a temporal feature vector; The construction of the risk prediction integrated model specifically involves performing tensor splicing on the joint graph feature matrix and the time series feature vector, and classifying the risk levels through a multi-layer perceptron to obtain a health risk prediction result.

3. The cloud-based elderly care information management system according to claim 1, characterized in that: In the privacy protection processing module, the federated feature aggregation specifically adopts an asynchronous gradient update strategy, generates feature embedding vectors through edge node local model training, and performs weighted aggregation in the cloud using a homomorphic encryption algorithm; The dynamic noise injection is specifically to dynamically adjust the Laplace noise intensity according to the data sensitivity. The noise injection formula is: Among them, ε t represents the noise intensity at time t, λ is the privacy budget coefficient, δ is the failure probability threshold, S t is the information entropy value of the current data slice.

4. The cloud-based elderly care information management system according to claim 1, characterized in that: The resource scheduling module specifically uses a deep reinforcement learning framework to make dynamic decisions based on the desensitized health feature data and the health risk prediction results, specifically including: defining state space, building a double-delayed deep deterministic policy gradient network, designing reward functions and online policy optimization; The defined state space specifically encodes the user health status, service resource inventory, environmental safety index and device availability into a multi-dimensional state vector; The dual-delayed deep deterministic policy gradient network is constructed, specifically using a master-slave dual critic architecture, generating resource configuration actions through a policy network, and achieving training stability through delayed updates of a target network; The design reward function is specifically to construct a composite reward function based on service response efficiency, resource utilization and risk mitigation degree: R=α·E r +β·U r +γ·R risk Among them, E r The response efficiency score is U r is the resource utilization rate, R risk is the unresolved risk weight, and α, β, and γ are dynamic adjustment coefficients.

5. The cloud-based elderly care information management system according to claim 4, characterized in that: In the comprehensive service module, the service recommendation method driven by the knowledge graph includes the following specific steps: constructing an ontology library in the field of elderly care, generating a dynamic knowledge graph, performing multimodal semantic matching, and outputting a personalized service chain; The construction of the elderly care ontology library specifically defines six entity types, including user portraits, medical resources, nursing services, social activities, safety equipment, and emergency strategies, and establishes 53 types of associations between entities; Generating a dynamic knowledge graph specifically involves incrementally updating the real-time collected environmental data, user behavior data, and service record data through a graph database to form a dynamic knowledge graph with a timestamp; The multimodal semantic matching is performed by adopting a cross-modal transformer model to jointly encode user voice requests, text work orders and sensor data to generate a unified semantic vector; The output personalized service chain specifically performs multi-hop reasoning in the knowledge graph based on semantic vectors to generate a comprehensive service sequence including medical service appointments, activity recommendations, and safety warning linkage.

6. The cloud-based elderly care information management system according to claim 1, characterized in that: In the health data collection module, the user's comprehensive health raw data specifically includes: heart rate variability data, blood oxygen saturation data, gait feature data, indoor positioning trajectory data, air quality index data and emergency call event record data; The edge computing node deploys a lightweight data preprocessing unit to specifically perform data format standardization, outlier filtering and feature dimension reduction operations.

7. The cloud-based elderly care information management system according to claim 2, characterized in that: The health risk prediction results output by the risk prediction module specifically include: disease onset probability value, fall risk index, loneliness tendency score and nutritional imbalance warning level; The disease onset probability value is calculated by analyzing 72 hours of continuous physiological data through a time series convolutional network, and the fall risk index is generated by combining gait characteristics and environmental obstacle distribution through graph neural network reasoning.

8. The cloud-based elderly care information management system according to claim 4, characterized in that: The resource scheduling module is configured with an elastic resource pool, which includes three types of schedulable resources: a mobile nursing robot cluster, an emergency medical supplies inventory, and an online expert consultation channel; The deep reinforcement learning framework performs a strategy update every 15 minutes, dynamically adjusts the allocation ratio of the three types of resources according to real-time demand changes, and verifies the feasibility of the scheduling plan through digital twin system simulation.

9. The cloud-based elderly care information management system according to claim 5, characterized in that: The integrated service module integrates a voice interaction interface, specifically adopts an end-to-end speech synthesis and recognition model, supports dialect adaptation and fuzzy command parsing; The voice interaction interface has a built-in anti-noise enhancement unit, which eliminates environmental noise interference through a generative adversarial network, ensuring that the command recognition accuracy rate is maintained above 90% under 60 decibel background noise.

Images